There is no federal ban on general-purpose Chinese AI models. That’s the fact, and a lot of this week’s coverage got it backwards.

What does exist are device-level bans, in place since early 2025, at Commerce, the Navy, NASA, the Pentagon, and the state governments of New York, Tennessee, and Virginia. That’s the entire scope of what’s actually active today. But per Fortune’s reporting on July 22, the mechanism to go far beyond that is already drafted and sitting in a drawer.

The mechanism exists, it just needs a signature

Fortune and ChinaTechNews reported two concrete instruments the administration has ready: an executive order that would require US companies hosting Chinese models to guarantee their security and accept liability for breaches, and a Commerce rule that would use supply-chain-security authority to restrict Chinese open-weight models specifically. Both were shelved months ago over concerns that clamping down would cost the US more in lost innovation than it gained in security. Moonshot AI’s Kimi K3 launch on July 16 revived that internal discussion.

Here’s what changes the risk calculus for any company: neither instrument requires a Congressional vote. An executive order gets signed. A Commerce rule gets published. There’s no mandatory public hearing, no vote, none of the notice period a legislative change would carry. The legality of building on a Chinese open model today rests on an administrative decision, not a stable legal framework.

The other half of the story: DeepSeek is already approved, just not published

Here’s a detail almost nobody is connecting to the Kimi K3 discussion. Per Reuters reporting via Yahoo Finance, an interagency committee already approved adding DeepSeek to the Commerce Entity List over a year ago. It’s still unpublished. That’s the longest gap between Entity List updates in more than a decade. This isn’t a government that decided not to act. It’s an action that’s approved and sitting on hold.

Treasury Secretary Scott Bessent was direct about the rationale on Fox Business on July 21: “If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them.” On the other side, Nvidia CEO Jensen Huang called Chinese open models “excellent” in an Axios interview and described backdoor fears as “a misconception,” since security controls can be applied to weights once they’re downloaded. Hugging Face CEO Clem Delangue downplayed the IP-theft claims and attributed China’s edge to strong research teams rather than copying anyone else’s work.

There’s no consensus inside the industry or inside the government. That disagreement, on its own, is useful information for any team currently deciding which model to build on.

There’s a third, slower track running in parallel: Congress. The bipartisan “No DeepSeek on Government Devices Act” (H.R.1121) and a broader proposal called the “No Adversarial AI Act,” which would cover all federal agencies, both remain pending votes. Neither is law yet. But that’s the path with actual advance notice, public hearings, and documented debate, exactly what the executive order and Commerce rule skip. Worth distinguishing the two tracks: the legislative one telegraphs itself months in advance, the administrative one doesn’t.

What this means for your model decision

This connects directly to something we already wrote about the half-life of business decisions: every decision you make about AI rests on conditions that decay over time, some faster than others. A model architecture decision usually has a half-life measured in months. A decision that depends on a regulatory instrument staying unsigned has a half-life nobody can calculate, because it depends on a signature, not a market cycle.

That doesn’t mean building on a Chinese open model is a bad decision today. It means it’s a decision with a specific, visible expiration condition: if that executive order or Entity List addition activates, your architecture goes from cost-optimized to out of compliance, potentially overnight, without the transition window a legislative change would give you.

How we handle this in discovery

In AI Maestro discovery, we treat model vendor dependency as a decision documented with its own explicit re-evaluation trigger, not an architecture choice made once and forgotten. If your company built on a Chinese model for cost reasons, the right question isn’t “is this legal today?” It’s “what specific event would make this illegal, and how close is that event to activating?” This week, the answer is: closer than most architecture teams assume.

Document the expiration condition on your model decision